[DPOTrainer] Fix DPO trainer + mistral + FA2#1290
Merged
Merged
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
kashif
approved these changes
Jan 30, 2024
Collaborator
|
thank you great help! |
lapp0
pushed a commit
to lapp0/trl
that referenced
this pull request
May 10, 2024
5 tasks
yxliu-TAMU
pushed a commit
to mincheolseong/ECEN743-GRPO-Project-Proposal
that referenced
this pull request
Apr 20, 2025
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes: #1217
Fixes: huggingface/transformers#26877
Fixes: #1266
Simply setting
use_cache=Falsecircumvents all issues with FA-2 + DPO + Mistral. in fact we should bypass that check since we are not in text generation mode when computing the loss function.use_cacheis retrieved from the model config by default which falls back always toTrue. The cache is not used anyway when purely computing the logits so this change is fully BCcc @kashif @vwxyzjn
To test that, I managed to repro the issue by adding
--attn_implementation "flash_attention_2"in the dpo shell script on a A100 machine, and I confirm this PR fixes it. Unfortunately our CI runners are not compatible with FA2 so we cannot add a slow test to test that